Thermal ablation destroys tumors by heating them with a needle-shaped probe. Large blood vessels next to the tumor carry heat away, which leaves a thin rim of living tumor along the vessel wall. This "heat sink" rim is a documented driver of local recurrence (Birrer et al.; Li et al.). Deep networks can learn to predict the patient-specific shape of the destroyed zone from pre-procedure CT. They are trained on zones that radiologists outline on follow-up scans. Those outlines are noisy in ways that matter most near vessels. Pre- and post-procedure scans are misaligned by up to a couple of millimetres, swelling blurs the boundary, and an annotator who expects complete destruction may draw the surviving rim as destroyed tissue. A large network will eventually memorize these errors.
The standard defenses against noisy labels make things worse here. They treat high-loss examples as suspect, and the true heat-sink indentations are exactly the high-loss, hard examples. This proposal makes a small, deliberately weak model of vessel cooling the judge of which label voxels to trust. That model is organized along the vessel tree, from trunk to branch, and it knows only that cooling grows with vessel size and fades with distance. A high-capacity segmentation network is then trained mostly on the voxels this judge agrees with. The bet is that a model too small to memorize, but shaped by the physics, can separate real cold spots from annotation errors where loss magnitude cannot.
Core Claim
Near vessels, label noise and true signal sit in the same place. Telling them apart therefore needs a prior about how the true signal depends on vessel geometry, not a prior about loss values or boundary smoothness. We claim that a capacity-restricted predictor built on an explicit multi-scale vessel tree can supply that prior. It is fit early in training, before the memorization phase described by Han et al. and Liu et al.. It is then used to weight the loss of an expressive 3D segmentation network, so that the network recovers heat-sink indentations better than either a noise-robust loss or simple vessel conditioning. The claim fails, and we will say so, if a plain 3D U-Net that receives the same vessel information and is trained with a robust loss performs just as well.
Where This Sits
The closest application work is Murthy et al. (also art_288b078f; Pre-operative lung ablation prediction using deep learning). They train a 3D U-Net on pre-procedure CT in a needle-centered coordinate frame and report a median Dice of 0.60 against 0.54 for the vendor's ellipsoid charts across 52 lung microwave ablations. They also report no volume bias, and their model visibly learns vessel indentations. This work motivates ours. It shows the signal is learnable, but it treats a single radiologist's 1-month outlines as clean and handles vessels only implicitly through U-Net pooling.
The closest noise-robust segmentation work is the hybrid architecture of Future Internet 2022. It estimates per-slice label quality from loss and adds a distance-transform shape regularizer. It constrains our design rather than supplying it. Its shape prior penalizes irregular concavities, which are what heat sinks produce. Its loss-based quality estimate shares the failure that Mercier et al. describe: once noise is memorized, loss stops separating clean from corrupted examples.
The noise-robust objectives we compare against come from classification. These are generalized cross-entropy (Sabuncu et al.) and SIGUA's selective unlearning (Han et al.). Both assume noise is unstructured with respect to the input. Ours is not.
From multi-scale pathology we borrow one structural idea. PTree-Net (Chen et al.) and DAS-MIL (Bontempo et al.) link coarse parent nodes to fine child nodes and enforce consistency between them. Shrivastava et al. show that flat concatenation of scales does not help. We apply the parent-child idea to a 3D vessel tree instead of a magnification pyramid.
The idea that a restricted-capacity model resists memorization follows Cheng et al. and the capacity analysis of Abbas et al..
Research Program
Vessel tree. We segment vessels in the pre-procedure CT and skeletonize them into a tree. Each node is a branch segment carrying its diameter, length, and position relative to the needle. Parent-to-child messages pass a coarse flow estimate down the tree. We use diameter as a proxy because absolute flow is unobserved, and we let a learned monotone function map node features to a cooling strength.
Judge model. For every voxel near the planned ablation zone, the judge predicts a correction to the vendor ellipsoid's probability of tissue destruction. The correction is a sum over nearby branches of cooling strength times a learned decay in distance. Monotonicity in diameter and in distance is enforced by construction. The judge has only a few hundred parameters, so it can express "big nearby vessels carve notches" but cannot memorize a particular annotator's boundary jitter.
Trust weights. The judge's output $g(v)$ at voxel $v$ is compared with the observed label $y(v)$. The training loss of the main 3D U-Net at that voxel is weighted by
$$w(v) = \exp\!\big(-\,|y(v) - g(v)|^2 / \tau(v)\big),$$
where $\tau(v)$ is a tolerance that grows with distance from the boundary of the judge's predicted zone. This growth reflects the fact that registration and swelling errors concentrate at the boundary. Disagreement deep inside a zone is penalized gently. Disagreement right next to a large vessel, where the judge is confident a notch belongs, is penalized strongly. Weights are kept away from zero so that hard but correct voxels still contribute.
Training schedule. The judge is fit first, on all labels, for a short schedule. It is then frozen and the main network is trained with the weights. In a second round, the judge is refit on high-weight voxels only, so that its estimate is not anchored to noise in the first-round labels. The main network sees the full CT and the vessel tree. The judge's only role is to decide trust, not to constrain the output shape.
Training data from simulation. Because the relationship between vessels and notches must be learned from very few clinical cases, we also pretrain on simulated ablations. We reimplement a simplified version of the vessel-aware bioheat solver of Audigier et al.. That solver couples a perfusion model of tissue with convective transport along large vessels, and the reimplementation runs on LiTS anatomy (Bilic et al.). Vessel and organ masks follow the pipeline of Chaitanya et al., and needle placements are sampled from their reinforcement-learning planner. Simulated zones are used as training targets, and their known ground truth also lets us measure exactly how well each method recovers notches under controlled corruption. We do not claim the simulator is accurate: Audigier et al. report 8.7 mm mean surface error on real ablations. We use it to study noise, not as truth about patients.
Baselines
The naive reference points are the vendor ellipsoid and a plain 3D U-Net trained with cross-entropy on raw labels. The latter follows Murthy et al., whose reported median Dice of 0.60, against 0.54 for the vendor model on their 52-ablation lung cohort, is the clinical number to beat. We will reproduce their applicator-centered input format exactly.
The strongest competitor on our specific question is a stronger version of that U-Net. It receives a vessel distance map and a diameter map as extra input channels and is trained with generalized cross-entropy (Sabuncu et al.). That paper reports 87.13% versus 81.88% for cross-entropy on CIFAR-10 with 40% noise. No reported result was found on Lacuna for generalized cross-entropy on ablation-zone segmentation. This baseline sees everything our method sees and has a standard noise defense, so it isolates whether the judge-based weighting adds anything.
We also run the label-quality and shape-prior hybrid of Future Internet 2022, which reports liver Dice of 56.99% against a 48.07% plain baseline on CHAOS under 75% synthetic boundary noise. We also run small-loss selection and SIGUA (Han et al.), adapted to voxels. All methods share the same backbone, input channels, simulated pretraining, and data splits. Each gets an equal hyperparameter search budget of the same number of trials, with selection on a held-out set labeled with the same kind of noise as training, never on clean labels.
Two ablations isolate the key ingredient. The first replaces the tree-structured judge with an unstructured small model that sees the same voxel-wise vessel features but has no parent-child messages. The second replaces the judge with a full-capacity network trained to convergence, which should memorize and lose its value as an arbiter.
To count as beating the strongest baseline, our method must improve Dice inside a 5 mm band around vessels wider than 3 mm and reduce error in notch depth. These gains must be larger than the seed-to-seed spread across five seeds, and overall Dice must not fall.
Why not just…?
Why not just feed vessel maps into the U-Net and use a robust loss? This is the real threat. If the network can learn the vessel-to-notch relationship directly, a bounded loss may already limit memorization enough. Our concern is specific to the "fill-in" noise, where an annotator draws the surviving rim as destroyed tissue. That noise is correlated with vessels, so a vessel-conditioned network can learn to reproduce the annotator's pattern as easily as the physics. A bounded loss does not help when the noise is systematic rather than random. The direct comparison on simulated data with controlled fill-in rates settles this question. If the gap is within noise across fill-in rates, the judge is unnecessary and we will report that.
Why not just use small-loss selection or early stopping? Heat-sink notches are hard voxels with high loss, so small-loss selection will discard them preferentially. Once memorization sets in, loss no longer flags noise (Mercier et al.). We will report the fraction of true-notch voxels each selector discards, alongside Dice in the vessel band.
Why not just train on simulator outputs, or use the simulator directly? The simulator's 8.7 mm clinical error (Audigier et al.) comes from generic tissue parameters, so its outputs carry their own systematic bias. Our judge uses only the direction of the physics, through its monotone constraints, and fits magnitudes from data. On clinical data we will compare simulator-only prediction and simulator-pretrained fine-tuning against our method.
Why not just add a shape prior? A smoothness prior is built to remove the concavities we are trying to preserve. We expect the hybrid of Future Internet 2022 to score well on overall Dice while underestimating notch depth. Notch-depth error will show this directly.
Proposed Experiments
Experiment 1: controlled noise on simulated ablations. On simulated LiTS ablations with known ground truth, we corrupt labels with three physically motivated noise types. The first is rigid misalignment up to 2 mm, which matches the registration cutoff used by Murthy et al.. The second is random dilation and erosion of the boundary, mimicking swelling. The third is fill-in of the perivascular rim at rates from 0 to 60% of cases. We report overall Dice, vessel-band Dice, notch-depth error, and volume bias.
Decisive check 1. The method is supported only if, at fill-in rates of 30% and above, it recovers notch depth substantially better than the vessel-conditioned robust-loss U-Net. Overall Dice should match that baseline at zero noise.
Decisive check 2. The unstructured-judge ablation tests whether the vessel tree matters. If the unstructured judge matches the tree judge, the multi-scale component is decoration and should be dropped.
Experiment 2: clinical retrospective cohort. We need a cohort comparable to the lung microwave ablation data of Murthy et al.; access is not assumed and must be negotiated. With such data, we measure Dice and volume bias against the 0.60 reference. We also have a radiologist blindly re-read the voxels our method down-weighted, to check that they are disproportionately errors in registration or rim annotation rather than true anatomy.
Experiment 3: memorization dynamics. Following the capacity lens of Abbas et al., we track the loss on corrupted voxels across training for each method. A working judge should keep that loss high late in training while the loss on clean notch voxels keeps falling.
References
- Radiofrequency ablation (RFA) in unresectable pancreatic adenocarcinoma: meta-analysis & systematic review
- Current evidence and strategies for preventing tumor recurrence following thermal ablation of papillary thyroid carcinoma
- SIGUA: Forgetting May Make Learning with Noisy Labels More Robust
- Robust early-learning: Hindering the memorization of noisy labels
- Pre-operative lung ablation prediction using deep learning
- A Hybrid Robust-Learning Architecture for Medical Image Segmentation with Noisy Labels
- Interpreting Deep Models through the Lens of Data
- Generalized Cross Entropy Loss for Training Deep Neural Networks with Noisy Labels
- Diagnose Like A Pathologist: Weakly-Supervised Pathologist-Tree Network for Slide-Level Immunohistochemical Scoring
- A Graph-Based Multi-Scale Approach With Knowledge Distillation for WSI Classification
- Deep Learning for Visual Recognition of Environmental Enteropathy and Celiac Disease
- Mitigating Memorization of Noisy Labels via Regularization between Representations
- Effective dimension of machine learning models
- Lattice Boltzmann Method for Fast Patient-Specific Simulation of Liver Tumor Ablation from CT Images
- The Liver Tumor Segmentation Benchmark (LiTS)
- Automatic planning of liver tumor thermal ablation using deep reinforcement learning